At what task difficulty does multi-agent decomposition become worth the coordination cost?
This explores when splitting a task across multiple agents actually pays off — at what point the gains from decomposition outrun the overhead, error propagation, and token cost that coordination introduces.
This reads the question as a threshold problem: there's a task-difficulty band where multi-agent decomposition earns its keep, and bands on either side where it doesn't. The corpus is surprisingly direct that the threshold is real and measurable — and that it's narrower than most people assume. The sharpest result comes from a study of 180 configurations finding that coordination *stops* helping above roughly 45% single-agent accuracy, while tool-coordination trade-offs actively *harm* the most complex tasks When does adding more agents actually help systems?. That's the counterintuitive part: the hardest tasks aren't where decomposition shines — they're where coordination overhead and error amplification (4–17× depending on topology) can sink you fastest.
Empirical analysis shows MAS performance gaps narrow with stronger models, with SAS outperforming in many cases. Three formal defect types—node-level bottlenecks, edge-level overwhelm, and path-level error propagation—explain when single agents win.
Research shows 80% of multi-agent performance variance comes from token budget, not coordination intelligence. LatentMAS and shared-KV-cache approaches offer ways to decouple performance gains from token costs.
Modular architectures with separate decomposer and solver models outperform monolithic LLMs, with decomposition ability transferring across domains while solving ability does not. The separation prevents planning-execution interference and produces more generalizable skills.
AgentsNet benchmark shows agents fail to coordinate strategies either by agreeing too late or adopting strategies without informing neighbors. Agents accept neighbor information without verification, enabling error propagation while remaining capable of detecting direct conflicts.
SLMs handle the repetitive, well-defined language tasks that constitute most agent work at 10–30× lower cost than LLMs, making heterogeneous architectures (SLMs by default, LLMs selective) the economically rational design pattern.
Research prompt for your LLMexpand ↓
Copy into ChatGPT or Claude to take this line of inquiry further — it asks the model to find newer work and re-test which earlier constraints still hold.
You are an AI systems researcher evaluating whether multi-agent decomposition remains cost-justified as model capability and orchestration infrastructure evolve. The question: *At what task difficulty does multi-agent decomposition become worth the coordination cost?*
What a curated library found — and when (dated claims, not current truth):
Findings span 2023–2026; treat as perishable:
- A 180-configuration study identified a sharp threshold: coordination *stops* helping above ~45% single-agent accuracy; above that, error amplification (4–17×) dominates (2025–2026).
- Tool-coordination trade-offs *actively harm* the most complex tasks, reversing the intuition that hardest problems benefit most from decomposition.
- Newer work (2026) finds single-agent LLMs outperform multi-agent systems on multi-hop reasoning *under equal thinking time*, suggesting the threshold may have risen or vanished.
- Recent coordination layers (semantic-aware fabrics, federation protocols, ~2025–2026) claim to reduce overhead; unclear whether they shift the 45% breakpoint.
Anchor papers (verify; mind their dates):
- arXiv:2308.00352 (2023-08, MetaGPT: earliest multi-agent framework benchmark)
- arXiv:2509.20175 (2025-09, AgentsNet: coordination overhead measurement)
- arXiv:2604.02460 (2026-04, single-agent LLM vs. multi-agent on multi-hop reasoning)
- arXiv:2512.08296 (2025-12, scaling laws for agent systems)
Your task:
(1) **RE-TEST THE 45% THRESHOLD.** Has this constraint held or shifted? Check whether newer orchestration (memory-sharing, cached reasoning, advanced routing) or frontier models (o1, o3-scale equivalents) have *relaxed* the error-amplification ceiling or pushed the breakpoint higher. Cite what changed it—or say plainly where it still holds.
(2) **Surface work contradicting the threshold claim.** Specifically, hunt for papers (last 6 months) showing multi-agent *gains* persist above 45% accuracy, or single-agent *failures* that multi-agent rescues, even on complex tasks.
(3) **Propose two research questions assuming the regime may have moved:** (a) Does *asynchronous* or *hierarchical* agent topology avoid error amplification where flat topologies failed? (b) Can learned routing or per-task decomposition policies overcome the static 45% rule?
Cite arXiv IDs; flag anything you cannot ground in a real paper.